Agent-Harness-Kit: Skill for Claude Code

Feature discovery

1. Read ../../../harness/playbooks/feature-discovery.md and follow it before creating or changing technical graph nodes.
2. Require approved project context. If it is missing, stale, or unapproved, route to first-run-discovery instead; do not run two interviews at once.
3. Activate without requiring the user to name this skill when the request introduces a new product capability and leaves consequential choices open. Also activate when graph planning or expansion exposes an unresolved functionality boundary or completion condition; an initial system brief is not blanket scope approval. Explicit brainstorming, ideation, "what if", "how could we", and feature exploration always qualify.
4. Do not activate for a bug fix, maintenance/refactor, dependency update, direct implementation of an approved brief, or a direct-trivial copy/style/static-content edit. The fast path precedes project discovery and creates no feature artifact.
5. Start by reflecting known project evidence and ask exactly one highest-leverage unanswered feature question. Never repeat the Harness welcome or ask the user to restate approved context.
6. Analyze the feature for consequential gaps before proposing implementation: intended and excluded actors, eligibility and permissions, entry/onboarding, the happy path, alternate paths, failure states, cancellation, recovery, empty states, data lifecycle, integrations, privacy/security, and observable success. Ask only the branches that apply. For authentication, for example, resolve email versus federated login, verification, forgotten-password recovery, account conflicts, provider failure, session/logout behavior, and who may access what.
7. Present two to four credible directions with tradeoffs and one recommendation. Do not silently select scope, UX, data, integration, architecture, or risk decisions for the user. State concrete "This task is complete only when..." conditions for the next bounded functionality; help the client resolve unclear rules with examples instead of deciding silently. A feature is not closed while a consequential journey, failure, recovery path, or acceptance outcome remains unresolved.
8. Persist the result in ../../../harness-state/features/FEATURE-<id>.md. Materialize FEATURE-BRIEF through the runtime template catalog only for that consumer, and keep it draft until the user explicitly approves the selected direction.
9. Do not add unapproved feature scope to PENDING.md or TASK-GRAPH.md before feature approval. If existing execution exposes an open condition, record the exact human blocker and keep the existing node pending/blocked with scope_status: needs-discovery; this records a stop, not new scope. After approval, update macro scope in pending state when applicable; when planning or implementation was requested, route through writing-plans before graph creation.
10. In hackathon mode, ask at most two cohesive feature questions before offering directions and converge on one demonstrable vertical slice; record deferred edge cases and shortcuts explicitly instead of pretending they were solved.